Gemma Scope 2: Helping the AI Safety Community Deepen Understanding of Complex Language Model Behavior

This post announces Gemma Scope 2, an expanded suite of interpretability tools for Gemma 3 models. It introduces full coverage at scale for models up to 27B parameters, skip-transcoders and cross-layer transcoders for deciphering multi-step computations, advanced training techniques like Matryoshka, and specific tools for analyzing chatbot behaviors such as jailbreaks and refusal mechanisms. The release involves storing ~110 PB of data and training over 1 trillion parameters.